But what this article truly discusses is not VVV’s short-term price surge, but a more fundamental question: Where will the value of AI platforms ultimately reside when model capabilities rapidly become commoditized?
The author’s core assessment is that leading AI labs such as OpenAI and Anthropic are trapped in an “equity structure trap”: their valuations rely on the assumption that model capabilities will remain scarce and command high premiums over the long term. However, China’s open-source models, low-cost training, open-weight ecosystems, and cloud-based deployment are rapidly driving down the price of model capabilities themselves. In other words, the most expensive segment of the AI industry may be becoming the hardest to sustain profitability in.
Within this framework, Venice is viewed by the author as an inverted structure: it does not train models, but rather leverages open-source model capabilities; it does not rely on centralized data storage, but emphasizes privacy and TEE proofs; it does not turn users into training data, but instead integrates users into the platform economy through mechanisms such as VVV staking, subscription burning, and DIEM compute rights. What the author truly aims to convey is that Venice is not merely a “tokenized AI application,” but an experiment in using tokens to reimagine consumer software relationships.
What matters most is not whether Venice can directly challenge OpenAI, but whether the AI market is splitting into two segments: one continuing to serve customers willing to pay for cutting-edge models and accept enterprise-grade compliance and data retention; and another shifting toward “good enough” open-source model capabilities, placing greater emphasis on privacy, censorship resistance, low cost, native agent access, and user ownership. If this split occurs, Venice’s opportunity lies not in winning the entire model war, but in becoming the inference layer and settlement rail for the open agent economy.
Therefore, this article presents a classic structural bullish argument: it is not merely betting on VVV’s price increase, but on the convergence of several trends—the commoditization of the model layer, open-source models catching up, the rise of agent-based payments, and the user-owned economy.
The risk lies precisely here—if progress on the open-source model slows, token burn fails to keep pace with growth, or Venice fails to genuinely cultivate user relationships, this narrative will be revalued. But at least for now, VVV’s market performance indicates that the market is beginning to pay a higher premium for this story of “the same demand, opposite economic model.”
The following is the original text:
These labs are spending hundreds of billions of dollars in an attempt to defend a moat that is evaporating in real time. GLM-5.1 has outperformed GPT-5.4 on the most challenging programming benchmarks—it is open source, licensed under MIT, and trained on Chinese hardware that the U.S. has sought to restrict. The cost of training cutting-edge capabilities has dropped by approximately 95% over the past eighteen months. Every dollar of OpenAI’s $852 billion valuation rests on the assumption that these changes don’t matter. But they do. And Venice is the only consumer-grade AI platform that will directly benefit economically when this reality is finally repriced by the market; even if such repricing never occurs, its investment thesis remains sound.
The central argument of the April article was that Venice holds a unique position in the agent economy. This assessment still holds—usage has tripled, the burn ledger has exceeded 42% of the genesis supply, DIEM has been repriced by 75% within six weeks, and the token price has more than doubled since I wrote that in-depth analysis.
But the "Seven Advantages" framework I proposed in April may have underestimated what is truly happening. Venice is not an AI company with a privacy label that happens to issue tokens. It is a new economic structure for consumer software: users are owners, the platform is the infrastructure, and value is measured not in equity but in computational power rights.
This structure is not a stack of features, but the only configuration capable of surviving the impending changes at the model layer. What the bubble is built upon, Venice stands in direct opposition to. The same market, the same demand, a completely opposite economic model. This is the mirror.
This is the point I didn’t clearly explain in April. Now I’m adding it.
OpenAI, Anthropic, and Together AI share a commonality unrelated to their products: their investors expect equity returns denominated in U.S. dollars, in the hundreds of billions of dollars, and demand these returns within an accelerated timeline.
It sounds ordinary until you follow this logic further.
OpenAI’s $852 billion valuation implies it will need to generate annual revenues of approximately $200 billion to $280 billion by 2030 to justify this multiple. The company currently earns $2 billion per month but posted a $13.5 billion loss in the first half of 2025; meanwhile, as inference costs surged fourfold to $8.4 billion, its adjusted gross margin declined from 40% to 33%. Compute and talent costs consume 75% of total revenue. Microsoft will also take an additional 20% before 2032. OpenAI projects its compute spending will reach $121 billion by 2028, with a loss of $85 billion that year alone, and profitability is not expected until after 2030.
Anthropic is caught in the same trap, just on a different scale. With a $380 billion valuation, a $30 billion ARR run rate, and projected training costs of $42 billion by 2029, Google committed $40 billion last month, and Amazon added $25 billion—but both are essentially circular cloud credits rather than true equity capital. The five major hyperscale cloud providers have collectively pledged $660 billion to $690 billion for AI infrastructure just in 2026. Goldman Sachs estimates cumulative spending from 2025 to 2027 will reach $1.4 trillion, roughly triple the spending from 2022 to 2024. Sam Altman has personally signed off on $1 trillion in AI deals, while OpenAI’s revenue stands at just $13 billion.
These are not ordinary companies. They are bets on sovereign-level infrastructure disguised as software firms. Their valuations require the model layer to remain perpetually expensive. But the reality is that the model layer is becoming increasingly cheaper.
Over the past 60 days, the relationship between AI capital expenditures and AI capability has broken down, as evidenced by the release of three open-weight models.
Z.ai's GLM-5.1, released on April 7, scored 58.4 on SWE-Bench Pro, surpassing GPT-5.4's 57.7 and Claude Opus 4.6's 57.3. It is open-sourced under the MIT license and trained entirely on Huawei Ascend chips, without using any NVIDIA hardware; meanwhile, Z.ai itself remains on the U.S. Entity List, prohibited from accessing H100s. Its API pricing is $1 per million tokens for input and $3.2 for output, making it 5 to 8 times cheaper than Claude Opus’s $5 / $25.
Kimi K2.6, released by Moonshot on April 20, became the top-ranked open-weight model on the Artificial Analysis Intelligence Index with a score of 54, compared to 57 for leading closed-source labs. It outperformed GPT-5.4 on agent tasks: achieving a HLE-with-tools score of 54.0, higher than GPT-5.4’s 52.1. It scored 80.2 on SWE-Bench Verified, nearly matching Claude Opus’s 80.8. Cloudflare priced it at $0.95 per input and $4 per output, making it approximately 15 times cheaper than Claude Opus under heavy workloads. The initial training cost for Kimi K2 was just $4.6 million.
Released on April 24, DeepSeek V4-Pro ranks second on the Intelligence Index, behind only Kimi K2.6, and outperforms all other models except the top three frontier proprietary labs. It is licensed under the MIT license. The training cost for DeepSeek V3 was $5.6 million.
Three Chinese labs, 60 days, all open source, all achieving or surpassing state-of-the-art performance on at least one major benchmark, priced 5 to 15 times cheaper, with one even running on sanctioned hardware. The capabilities that supported OpenAI’s valuation in 2024 are now freely downloadable from Hugging Face, deployable on rented hardware, and continue to improve every quarter.
This is not the so-called “China AI moment.” Structural arbitrage at the model layer is happening in real time. A scholarly paper from March 2026 directly stated: “Pre-training scale has become decoupled from cutting-edge AI capabilities.” The share of Chinese open-source models in global usage has grown from 1.2% in 2025 to 30%. Apple is evaluating DeepSeek, Qwen, and Doubao for integration into iOS 27. AWS, Azure, and Google Cloud all offer deployment options for DeepSeek. Today, 80% of startups seeking VC funding are built on open-source models. Meta’s Llama series was intentionally released to drive commoditization at the model layer—when a company valued at $1.6 trillion is the most aggressive price reducer in your market, it’s clear where margins will flow.
Every dollar of OpenAI’s $852 billion valuation assumes these developments are irrelevant. It assumes enterprise customers will indefinitely pay premium prices for token-based capabilities, even when GLM-5.1 offers similar performance at one-eighth the cost; it assumes Kimi K2.6’s open weights are insignificant; it assumes DeepSeek’s offerings at less than 3% of the price of frontier models don’t matter. It assumes these labs can simultaneously achieve tenfold revenue growth and expanding profit margins in a market where competitors offer their products for free.
Jai Das of Sapphire Ventures has referred to OpenAI as the “Netscape of the AI era.” Mark Zuckerberg has also publicly acknowledged the existence of an AI bubble dynamic. In March, the Pentagon officially classified Anthropic as a supply chain risk because Anthropic refused to allow Claude to be used for mass surveillance or autonomous weapons; in contrast, OpenAI and Google signed “all lawful uses” agreements to avoid the same fate. Centralized AI companies are subject to government coercion, and their architectures cannot reject such coercion—Venice’s architecture can.
These labs are not unaware of the issues—they simply cannot pivot. The investors who wrote checks based on an $852 billion valuation did not buy into a future where the model becomes commoditized. They bought into a future where the model maintains a premium valuation at all times. These are two entirely different companies, and for the latter to truly materialize, it must first write down the valuation of the former.
This is the trap. The issue isn’t with the rejection mechanism stack or the logging architecture. The real problem is that the only investors who can tolerate Venice’s economic structure are those who already hold VVV.
From here, this argument no longer requires a bubble burst to hold true.
Assume these labs barely survive. Assume GPT-6 remains the best in its class, Claude Opus 5 continues to lead in reasoning, and Gemini maintains its frontier in multimodality. Assume enterprise contracts last long enough for these companies to secure refinancing and weather their valuation pressures.
This isn't important either. The market will split.
Cutting-edge intelligence accounts for only a small fraction of total inference demand. The vast majority of real-world workloads—programming assistance, writing, analysis, image generation, video, agent execution, customer service, research, summarization—reached "good enough" levels months ago. GLM-5.1’s coding capability in production is now comparable to GPT-5.4. Kimi K2.6’s agent execution capability is on par with Claude Opus 4.6. DeepSeek’s general reasoning ability is essentially on par with any model outside the absolute top of the leaderboard. For 80% of real-world needs, the open-weight ecosystem is already sufficient—and improving every quarter.
These requirements demand not greater intelligence, but intelligent attributes that labs are structurally unable to provide: privacy, uncensored outputs, no account required, no logging, native agent access, predictable costs, and user ownership. Labs serve a small subset of high-end customers willing to pay enterprise prices and accept surveillance. Venice serves everyone else—and that’s precisely the larger, faster-growing half of the market.
In a bull market scenario: these labs collapse, and Venice takes over the entire market. In the base case scenario: the market splits, with Venice holding the larger side. Even in a bear market scenario—where these labs maintain long-term dominance in cutting-edge capabilities and no repricing events occur—Venice remains one of the few consumer-grade AI platforms capable of serving the 80% of inference demand that doesn’t require state-of-the-art capabilities and cannot accept the labs’ business models.
This argument does not require a crash to occur. It only requires the open-source curve to continue along the path it has already taken.
Why did Venice capture this larger half of the market? Not because it was destined to win all. It might, but the structural answer is simpler than that.
Venice is the only consumer-facing AI platform that allows users to own the equity of the services they use. Stake VVV to earn yields and gain lifetime Pro access. Lock sVVV to mint DIEM, securing a permanent compute equity stake that appreciates as inference costs become commoditized. Every paying user fuels a burn cycle that compounds and enhances the positions of all other users. This isn’t just a feature—it’s an entirely different relationship between consumers and products, one that Big AI cannot offer, as their equity structures cannot accommodate “users as owners.”
Now consider what users truly need that labs cannot provide. Privacy is not a policy—it’s verifiable TEE proofs, zero retention, and an architecture where nothing can be seized. For 99% of intelligent use cases that don’t require filtering through corporate brand safety committees, uncensored output is essential. Open-source frontier models can be deployed within days of release, because Venice doesn’t need to defend a moat that forces the model layer to remain perpetually expensive. Agent-native access—autonomous API keys, x402 wallet payments, zero human intervention—because the agents being deployed today simply cannot work with anything else.
Each of these forces is independently intensifying. As data breaches rise and regulations tighten, demand for privacy is growing. As users grow increasingly frustrated with "brand-safe AI products" that routinely block everyday tasks, demand for censorship resistance is growing. Open source is narrowing the "good enough" gap every quarter. The share of agents in total reasoning demand is doubling. None of these forces point to a lab—they all point to Venice.
A platform built on the opposite of every bubble assumption—many of its features seem random until you see the overall shape.
Zero training cost. Venice has never spent a dollar training a model. Every release from Llama, Qwen, Mistral, GLM, DeepSeek, and Kimi is a free upgrade. Those labs have spent hundreds of billions of dollars trying to maintain a lead measured in months. Venice spends nothing and rides directly on the curve they pay to push. When GLM-5.1 was released at one-eighth the price of Claude, it was a margin expansion event for Venice—but a survival threat to companies trying to charge premium prices for equivalent capabilities.
Zero data retention. At laboratories, privacy is a policy commitment; at Venice, privacy is a mathematical structure. OpenAI Enterprise does not default to using customer data for model training, and customers can set retention windows—but during inference, prompts still pass through OpenAI’s servers and may be accessed by authorized personnel for abuse investigations, support, and legal matters. Policies can change. Vendors can also be compromised—in November 2025, Mixpanel leaked API customer names, emails, and organization IDs via SMS phishing. Runtime data can also be exfiltrated through novel vulnerabilities—Check Point disclosed a ChatGPT vulnerability in March that silently leaked conversation content via a DNS side channel. Even with contracts mandating zero retention, the architecture remains trust-based. Venice’s TEE attestation transforms privacy guarantees into cryptographic guarantees. Secure enclaves process prompts, return results, prove execution, then discard inputs. Venice cannot see your data because the architecture forbids it. This is not a privacy moat—it is a balance sheet that strengthens as data regulations tighten.
The token appreciates in value through usage-based mechanics. Every paid request buys and burns VVV on the open market. Tiered subscription burns scale with revenue growth: Pro ~$2, Pro+ ~$5, Max ~$10. Over the past 18 months, emissions have been reduced five times, with another halving planned before midsummer. 42% of the genesis supply has already been burned. No allocations go toward investor returns, as there are no investors. Every dollar of revenue is compounded back into the assets owned by stakers.
Users are an asset class, not a product. This is a point that no one has clearly articulated. On centralized platforms, users generate data; that data becomes training input, and that training input becomes the platform’s moat. Users are the product. On Venice, users consume tokens through staking, subscriptions, and paying for inference fees—tokens are burned, thereby increasing the value of every holder’s position. Users are an asset. The economic vector is completely opposite to that of virtually every other consumer software business in the world.
DIEM is a fixed-income instrument powered by reasoning capacity. One staked DIEM equals a daily-renewing $1 credit, valid indefinitely. It can be traded on Aerodrome or redeemed by burning to unlock the original sVVV stake. During the lock-up period, it earns approximately 80% of the standard VVV staking yield. This is not a regular token, but a fixed-income instrument backed by AI infrastructure. As underlying compute power becomes commoditized, each DIEM can purchase more reasoning capacity annually while its nominal claim remains unchanged. The lab is issuing equity based on an asset that is depreciating; Venice is issuing perpetual claims on an asset that is continuously appreciating.
Put together, what you get is not “an AI company with a crypto flavor.” You get an entirely new form of consumer software: every economic relationship between users and the platform is mediated by assets owned, priced, traded, and profited from by the users themselves. And whether or not those labs survive, these properties hold true. They are not a bet on collapse, but a structural advantage that compounds in any macro environment.
The agent economy is coming, and the timing coincides precisely with these labs running out of funding runway.
Coinbase Agentic Wallets have processed over 165 million transactions on x402. Google AP2 has launched with over 60 partners. Visa has released the Trusted Agent Protocol. Mastercard has invested $1.8 billion in stablecoin infrastructure—the largest stablecoin transaction ever. Coinbase launched Agent.market in April, with 69,000 active agents trading on the platform. McKinsey estimates that by 2030, agent-mediated consumer commerce will reach $3 trillion to $5 trillion.
Each of these agents requires a reasoning service provider. But they cannot be used in serious scenarios with OpenAI or Anthropic. The lab’s compliance architecture requires KYC; their revenue models require logging; and their content policies require rejection. Agents cannot fill out registration forms, enter CVVs, or agree to terms of service that may change next quarter. Coinbase’s CEO put it plainly: AI agents cannot meet KYC requirements or use traditional banking systems.
Thus, as China’s open-weight models arbitrage beneath the core businesses of these labs, the most important new demand category in AI infrastructure—autonomous agents—is structurally incompatible with their architectures. Agents reinforce market fragmentation: high-end demand remains at the top, while everything else moves toward agent-native.
Venice serves both ends of this transaction. The autonomous API key flow is live—smart agent staking VVV, signing tokens, minting keys, and paying with DIEM, all without human intervention. x402 wallet payments are now live on all paid endpoints. A single credential grants access to JSON-RPC across 11 chains. Every Eliza, Fleek, OpenClaw, Hermes, and NanoClaw agent is ready to use out of the box. The agents being deployed today run on Venice’s infrastructure because no other option can simultaneously offer permissionless access, privacy, censorship resistance, and native agent support.
When the commercial scale of agent intermediaries reaches the trillions of dollars predicted by McKinsey, and those labs hit the walls built into their equity structures—whether or not they actually do—Venice has already become the reasoning layer of this economy.
Something that is compounding
The arguments for April are no longer speculative. On April 7, daily usage reached 50 billion tokens and 1 million images. GLM-5.1, Kimi K2.6, and DeepSeek V4 all launched on Venice within days of their release, with privacy contracts remaining unchanged. DIEM’s execution discount has been repriced from 57% in early March to approximately 32% today—the market is repricing reliability, not incremental utility. Once the discount falls below 20%, DIEM will cross $1,500 purely through mechanical mathematics. Staking inflows have exceeded $15 million. Over 32 million VVV tokens have been staked, locking up approximately 70% of the circulating supply. The tiered subscription burn mechanism went live in April and is generating significant monthly burns; at current rates, even without the next emission reduction, VVV will turn net deflationary in Q3.
Every judgment in the April article has either been compounded or become even clearer. None have been weakened.
The April article stated that Venice is the only platform combining seven specific advantages—a judgment that still holds. However, what I didn’t clearly explain at the time was why: these seven advantages are not a stack of features, but rather the natural form of a consumer software company that doesn’t need to meet venture capital equity return requirements. Those venture capitalists are investing in equity based on assets that are about to be commoditized.
There are two possible paths for this market. The first is that these labs are overwhelmed by their own equity structures, and Venice takes over the entire technology stack. The second is market fragmentation—labs retain a small segment of high-end demand willing to pay enterprise prices and accept monitoring, while Venice owns everything else: the larger, faster-growing half of the market, where “good enough” intelligence combines with privacy, uncensored outputs, agent-native access, and user ownership.
The endpoint of both paths is Venice becoming the reasoning layer for an open agent economy. This argument does not require a bubble to burst; it merely requires the open-source curve to continue along the trajectory it has already set—fact is, it does so every quarter, faster than the market updates its models.
Venice is built on this bet. Three months ago, when I made this call at $2, no one was listening. A month ago, when the price reached $8, people started to pay attention. Now, at $18, the market still hasn’t fully grasped this structural argument—the part that hasn’t been priced in is what happens when these two scenarios ultimately converge on the same answer.
The bubble is built on the assumption that the model layer will maintain a high premium indefinitely. Venice’s compounding is based on the trend of the model layer moving toward being free. Whether the bubble bursts suddenly or deflates slowly, the outcome of this trade remains the same.
The same market. Opposite economic models.
The lab cannot keep up. Miners cannot capture users. The protocol is being handed over to the foundation. Value will ultimately concentrate in a few places, as always: the brands people choose, the tracks on which agents operate, and the currencies they use to price things.
Venice is building its brand, operational infrastructure, and issuing currency.
The next chapter is not a celebration. The real question is: Will the structural argument made in April’s article be repriced as venture-backed comparable companies run out of options, or will it be repriced as the market naturally splits around them?
Based on current evidence, both events are proceeding on schedule.
Not investment advice. Please conduct your own research.
[Original link]
律动 BlockBeats